Argus HQ — AI Disclosure
Last updated: July 17, 2026 Version: 1.0
This page is the feature-by-feature map of where Argus HQ uses an AI (large language model) and which model, and where it does not. It is a companion to the AI Disclaimer, which states the liability and verification rules that apply to all AI-generated content. If anything on this page conflicts with the AI Disclaimer or the Terms of Service, the AI Disclaimer and Terms control.
We keep this page in sync with the code that actually ships. If a feature below is not marked "AI-assisted," treat it as deterministic — sourced directly from the public record with no model in the loop.
What is NOT AI-generated
- The underlying source records. Every FDA Warning Letter, recall notice, approval record, or Form 483 inspection Argus HQ shows you is ingested verbatim from the public government source (fda.gov and related feeds). Argus HQ does not write, edit, or paraphrase the source record itself — only the commentary layered on top, as described below.
- Watchlist matching. Whether a record matches your configured watchlist (industries, CDMOs, competitors) is deterministic rule-based matching, not a model judgment call.
- Trend pulse and repeat-offender flagging. Warning-letter volume, month-over-month deltas, issuing-office breakdowns, and 24-month repeat-offender history are computed directly from the ingested record set — counting and lookups, not AI inference.
What IS AI-assisted, and with which model
All AI features below currently use Anthropic Claude. Every one of them is descriptive (states what the record says), never prescriptive (never tells you what to do), and every one has a deterministic, brand-voice-safe fallback that ships instead whenever the AI call fails, times out, or drifts from our content rules — so a customer never receives an unreviewed or off-policy AI output.
| Feature | Where you see it | Model | Notes |
|---|---|---|---|
| Alert summary | Daily brief email, dashboard | Claude Haiku | 2-3 sentence summary of a single regulatory record. |
| Warning Letter page analysis | Public /fda/warning-letter/[slug] pages | Claude Haiku | ~250-word longer-form analysis for the public SEO page version of a letter. |
| Stakeholder memo | Daily brief email (attached to each letter) | Claude Haiku | A one-sentence business-implication line plus 2-3 discussion-starter questions. |
| Inverse Argus generated pages | Public /fda/warning-letter, /fda/recall, /fda/approval, /fda/inspection-483, and company-rollup pages | Claude Sonnet 5 | Generates the page content from the source record, then is graded by a second Claude pass before anything publishes; failing pages are fixed or quarantined, never shipped ungraded. |
| Landing-page chat widget | argushq.ai marketing site, bottom-right chat | Claude Sonnet 5 | Grounded only in real, ingested Warning Letters via read-only lookup tools — instructed never to invent a company or citation. |
| Support-inbox draft replies | support@argushq.ai auto-triage | Claude Sonnet 5 | Classifies and, for five low-risk categories only (billing, feature question, delivery issue, unsubscribe, out-of-office), drafts a reply that may auto-send. The other five categories (refund, cancellation, legal, competitive, enterprise pitch) are always routed to a human. |
| Case-study drafts | Internal — drafted from a customer's own reply, reviewed by a human before anything publishes | Claude Sonnet 5 | Never auto-publishes; a person always reviews before it becomes customer-facing marketing copy. |
Guardrails that apply to every AI feature above
- No training on your data. Argus HQ does not use Customer Data to train or fine-tune any AI model, and Anthropic is contractually barred from training on data submitted through their API. See Section 6.3 of the Terms of Service.
- Brand-voice and accuracy linting. Every AI output is checked against a banned-phrase and drift filter before it ships; anything that fails is replaced with a deterministic, non-AI fallback derived directly from the source record.
- No invented facts. Prompts explicitly instruct the model never to invent a citation, date, company, or Warning Letter that is not in the ingested record set.
- Always verify against the source. Every AI-generated summary links back to the original government source URL. Per the AI Disclaimer, you should verify anything material to a business, legal, or regulatory decision against that source before acting on it.
Contact
Questions about this page or how a specific summary was produced: support@argushq.ai. Digital Empire Holdings LLC, 30 N Gould St Ste N, Sheridan, WY 82801.